John Harker

Accelerated Flash performance with Hitachi Storage Virtualization Operating System

Blog Post created by John Harker Employee on Dec 10, 2015


One criticism we hear about the VSP family is that it isn’t ‘designed from the ground up for flash’.   Which is odd when said about a system that at the high end is capable of producing over 4M IOPS sustained performance.   Over time, everything in the current generation VSP systems hardware and software has been optimized for the utmost in performance, including for flash. While at the same time providing ultimate reliability, protection and online serviceability.

With the VSP family the answer to flash optimization not just in the hardware, it is also in the Hitachi Storage Virtualization Operating System (SVOS).   Hitachi storage controllers are complex clusters of processors sharing various interconnect, cache, IO, and media devices. SVOS orchestrates connections and data transfers throughout the system. SVOS clustered o/s design provides massive parallelism for a non-blocking architecture for extreme throughput capabilities.  SVOS includes significant flash-focused engineering, patented, with over 30 fundamental changes to enable higher thread-count processing and faster movement of data.   SVOS is architected and optimized for maximum flash throughput throughout.

With flash storage, high performance design starts with high performance data paths between flash devices and servers.  Why solve the I/O bottleneck in the device only to push the bottleneck up into the storage controller?  The significant determinate of performance is the speed of interconnects and number of moves/copies of the data in the path.   Within Hitachi storage systems interconnects and RAID calculations are accomplished with a mix of custom silicon logic machines and multiple parallel PCIe busses, achieving over 830GB/s internal bandwidth and sub-millisecond response times.  So hardware bandwidth is high and latency is low for extreme performance . 

But extra internal data copying steps can still kill performance.  SVOS is optimized with flash to minimize them – for flash it doesn’t copy into a buffer or cache and then out again if it can be copied directly. For example one of these optimizations is unique cache bypass technology.  With HDS storage systems and traditional media, writes go through copies to cache, for subsequent high performance reads with data consistency. With (slower) disk systems this extra copy on the write doesn’t affect performance, but with faster flash devices it does.   With FMD or FMD2 media SVOS avoids this extra copy into cache with flash reads and writes using patented “express” I/O processing and optimization with caching algorithms and cache bypass optimized for flash

Also within SVOS there is media specific tuning of cache and I/O handling for specific devices. Logic for random read miss/write miss throughput, usually the major part of backend I/O processing, is all optimized for flash.

Within the system Dynamic Provisioning provides wide striping that automates performance tuning, self optimizes performance and capacity allocations, eliminates hotspots and improves wear leveling.  Dyamic Tiering has flash specific read/write IO profiling, modified tiering algorithms and buffer handling, and the new Active Flash feature that enables real-time elevation to flash of suddenly active data on lower Dynamic Tiering tiers.

So at the end of the day, as compared to an evolved SVOS and Hitachi Storage, there are no performance advantages of systems designed only for flash.   And lots of advantages with a mature fully featured system like the VSP family.